A parallel algorithm for scientific visualization

Scientific visualization aims to reveal relevant information from multi-dimensional datasets, e.g., scalar or vector fields, by carving out interesting structures and displaying them on the screen. Due to the algorithmic complexity and the large size of typical datasets (up to 512 ^-3 samples), interactive visualization requires either a workstation network, supercomputers or special purpose hardware. We describe a parallel implementation of a visualization tool for scalar fields on a workstation network. Challenges are to overcome the limited transfer bandwidth of the network, to establish an even workload and to avoid stall conditions in a multiuser environment, in which any machine might temporarily be allocated to other tasks. We alleviate the first problem by the use of BTC (block truncation coding), an image compression method which allows image fragments computed by the different workstations to be compressed and decompressed very quickly. Extended to 3D and applied to the dataset itself, it also speeds lip the sequential part of the visualization process. A modified self-scheduling scheme, which assigns work packages redundantly, eliminates delays caused by temporarily deallocated machines. All communication, synchronization and scheduling is done using basic constructs of PVM (Parallel Virtual Machine). Image quality and processing speed are illustrated by an example from medical imaging